A study co-authored by Dwarkesh Patel and Jerry Hahn found that over a six-year period, pre-training progress derived more than three times more from data improvements than model improvements